def feed(self, input_data, time_steps=1):
"""calculate result for input vector and cache intermediate results"""
# if number of cache exceeds <tim_steps>, remove first cache
if time_steps <= len(self.caches):
self.first_cache.successor.remove()
self.caches.pop(0)
# create new cache
cache = LayerCache()
self.last_cache.insert_before(cache)
self.caches.append(cache)
# cache input vector
cache.input_values = input_data
# calculate and cache output vector
cache.output_values = super(CachingNeuralLayer, self).feed(input_data)
return cache.output_values
def __init__(self, in_size, out_size, activation_fn, activation_fn_deriv):
"""initialize layer (compare class NeuralLayer)"""
# initialize cache chain (self.first_cache and self.last_cache are dummy caches)
self.first_cache = LayerCache()
self.first_cache.is_first_cache = True
self.last_cache = LayerCache()
self.last_cache.is_last_cache = True
self.first_cache.successor = self.last_cache
self.last_cache.predecessor = self.first_cache
self.caches = []
# invoke super constructor
super(CachingNeuralLayer, self).__init__(in_size, out_size, activation_fn, activation_fn_deriv)
class CachingNeuralLayer(NeuralLayer):
"""NeuralLayer that supports caching its inputs, results and losses"""
def __init__(self, in_size, out_size, activation_fn, activation_fn_deriv):
"""initialize layer (compare class NeuralLayer)"""
# initialize cache chain (self.first_cache and self.last_cache are dummy caches)
self.first_cache = LayerCache()
self.first_cache.is_first_cache = True
self.last_cache = LayerCache()
self.last_cache.is_last_cache = True
self.first_cache.successor = self.last_cache
self.last_cache.predecessor = self.first_cache
self.caches = []
# invoke super constructor
super(CachingNeuralLayer, self).__init__(in_size, out_size, activation_fn, activation_fn_deriv)
def feed(self, input_data, time_steps=1):
"""calculate result for input vector and cache intermediate results"""
# if number of cache exceeds <tim_steps>, remove first cache
if time_steps <= len(self.caches):
self.first_cache.successor.remove()
self.caches.pop(0)
# create new cache
cache = LayerCache()
self.last_cache.insert_before(cache)
self.caches.append(cache)
# cache input vector
cache.input_values = input_data
# calculate and cache output vector
cache.output_values = super(CachingNeuralLayer, self).feed(input_data)
return cache.output_values
def get_deltas(self, last_deltas, last_weights):
return [super(CachingNeuralLayer, self).get_delta(cache.output_values, last_delta, last_weights)
for cache, last_delta in zip(self.caches, last_deltas)]
def learn(self, result, delta, learning_rate):
#Logger.log("(expected delta: " + str(np.shape(self.caches[0].output_values)) + ")")
#Logger.debug("delta: " + str(delta) + "\nresult:" + str(result))
super(CachingNeuralLayer, self).learn(result, delta, learning_rate)
